Keith Rabois | 20VC Summary

1. Key Themes

The Great Late-Stage Valuation Debate: When Price Becomes the Only Risk

The fundamental tension of 2025 venture investing centers on valuation risk at unprecedented scales. When discussing SpaceX's $800 billion secondary at 40X+ revenue, Roelof Botha crystallized the core dilemma: "You're not running any other kind of meaningful risk, except valuation risk. So you can cry like a baby when valuation risk bites you in the ass. At one level, not only is it baked in the business model, it arguably is the business model." [00:06:43]

The panel revealed a remarkable shift in venture psychology - the complete abandonment of concerns about down rounds. Keith Rabois observed: "I think the meta point for me, the change to venture and founders is even today on X...you'll have advice, don't raise around at too high price, you'll come to regret it, do a fair price, don't take the highest price, blah blah blah. I think that advice is as sound as it ever was, but I think it has an audience of n equals zero. I don't think anyone cares." [00:07:48]

This extends to extraordinary pricing mechanics where companies like Harvey can raise $160 million at $8 billion with less than 2% dilution, forcing investors to pay "three years ahead" for public market multiples. The acceptance of this represents a fundamental recalibration of venture economics, where momentum and positioning trump traditional valuation discipline.

AI Companies Operating Under Different Physics Than SaaS Ever Did

The panel identified that AI companies are experiencing growth patterns fundamentally different from the SaaS era, which changes all traditional valuation frameworks. Harvey's trajectory from $10M to $50M to $150M ARR demonstrates 10X then 3X growth - a deceleration pattern that would terrify SaaS investors but still represents extraordinary performance.

Roelof Botha explained the analytical framework: "In SaaS, you could use this following rule of thumb with a high degree of accuracy. The growth rate for any year was roughly 85% of the prior years growth. So if it's growing at 100, it probably go at 85. If it's growing at 85, it probably go in the high 60s...The thing here is, instead of saying that best in class 3X growth, you're seeing best in class 10X growth. If that only declines slowly, if it goes 10X to 6X to 3X, these are amazing companies and any price can be paid." [00:48:26]

However, Keith Rabois introduced the instability thesis: "We haven't even begun to see deep reasoning in B to B apps. We haven't even started...You can't wait five minutes for Sierra or Decagon or Finn to give you an answer. No one's going to wait five minutes for that answer. When that goes from five minutes to five seconds or better yet, a second, just like we thought Jasper was amazing, we haven't even begun to imagine what B to B will do with AI when it can do deep reasoning in a second." [00:58:56]

The Coming IPO Wave: Netflix Eating Hollywood as Blueprint for Tech Eating Everything

The potential 2026-2027 IPO cohort represents one of the largest wealth creation events in venture history. The math is staggering: "800 billion for SpaceX, 400 billion for Anthropic and 200 billion for Databricks. That's 1.4 trillion of market cap. Let's say VC's on an average, little under half of that plus or minus, that's 700 billion of money returned to VC," calculated Roelof Botha. [00:09:26]

But more importantly, this represents tech's complete consumption of traditional industries. Roelof Botha provided historical context: "Netflix won, they ate the media industry, their market cap is $470 billion. The biggest studio is sub 200 billion dollars...The zoom out comment is, the outsider, the tech backed, the venture backed company has basically a far better business model than any of the studios and is eventually gone up, as you're seeing it here, start to eat them." [00:19:30]

This pattern extends beyond media: "They swallowed advertising completely, they swallowed a lot of entertainment, not all of it, they swallowed a good slug of retail, probably bank is gonna go that way, and you know, at some point maybe auto. It's just amazing over an elongated period of time, just the power of digital software enabled technology in conjunction with the internet to just, you know, each entirely destroy and consume other businesses." [00:30:29]

2. Contrarian Perspectives

Tiger Global's $2.2B Fund Represents Love of the Game, Not Capitulation

While the media narrative frames Tiger's downsizing from peak 2021 levels as retreat, the panel identified it as sophisticated market timing. Most contrarian was Roelof Botha's take: "It speaks to a love of the game that he says, I'm going to swallow my pride, because I don't think he's going to say the 21 fund was success or the right idea...I give him huge credit for not going, and now I'm just going to take my $2 billion, $5 billion, and stay in my house in Palm Beach or whatever. I'm going to get back in the game where there's a smaller fund, take the little bit of knots from the knockers." [00:13:01]

The truly contrarian element: a 20% GP commit at $2.2 billion scale. Keith Rabois noted: "20% is at the limit where it's almost not worth doing. You might as well just direct invest. You're just not getting enough leverage on your money...For folks who might be founders or others, that's a lot for a fund, right? I think the average growth fund is about 1%." [00:35:13]

This represents conviction pricing - Tiger isn't retreating, they're concentrating with real capital at risk, having done only 9 deals in 2025 versus 100+ in 2021.

The "TAM Trap" Applies to Old Companies, Not AI - But AI May Be Unstable

Harry Stebbings pushed back on conservative TAM thinking: "Every time we try to do too negative on TAM, it bites us...I believe in it, I'm living it with my portfolio." But he added a critical caveat about Harvey's legal AI dominance: "You need to believe they subsume all of those [vertical competitors] and be able to expand broader. I think you have to believe if you're going to see this being attractive from you...You also have to see market dominance. You also have to be the winner." [00:42:37]

Keith Rabois introduced the most contrarian take on AI stability: "All of these apps, we could say, oh, they'll just rotate out their LLMs, but it might not be that simple. At least could all become obsolete in an area of infinite deep reasoning...As investors, we have to imagine a world where that all goes away in the next two years. And it's not stable. And all of our investments could go to zero." [00:55:04]

This creates a paradox: pay premium prices for companies that might be rendered obsolete by model improvements, or miss the biggest wave in tech history.

LLMs Are Commodities - But That Doesn't Make Them Bad Businesses

Marc Benioff's "LLMs are commodities" comment sparked deep analysis. While the panel agreed on easy switching between models (unlike cloud providers), Roelof Botha rejected the business implications: "If you look at the hard disk market, it was a commodity. It was a brutally competitive for a while. It remained a commodity even after it consolidated to three, but they were able to extract enough profits from it...I think it's totally plausible that even though there is a fair degree of switching possible between models, you still will evolve to an industry structure where those model companies make decent money." [01:04:01]

The contrarian insight: commodity infrastructure at scale with high fixed costs naturally consolidates to oligopoly with acceptable returns. The challenge is surviving to consolidation.

3. Companies Identified

Harvey

Description: AI-powered legal software company
Why Mentioned: Exemplifies extreme late-stage AI valuations with operational excellence
Metrics: $150M ARR growing 300% YoY, 98% GDR, 168% NDR, raised $160M at $8B valuation (less than 2% dilution)

Keith Rabois provided the bull case: "If Harvey went from 50 to 150 this year...and if the last quarter, last four months, if the growth's consistent, so you can see a path to 400 next year, this is the same multiple we just talked about for anthropic, 20X...If this growth is durable...if it really has 170% NRR and 98% logo retention, and it's accelerating at 150, this is just the bet you do." [00:45:04]

Harry Stebbings challenged with competitive dynamics: "The challenge to that, there is LaGoura, 40 million ARR also, a lot less, but they're doing 10X euro near growth on a smaller base...LaGoura's European progress can't be underestimated. And really the analysis from investors today is that Harvey won the US and LaGoura winning Europe." [00:43:45]

Anthropic

Description: AI safety-focused frontier model company
Why Mentioned: Potential 2026 IPO candidate with extraordinary growth at massive scale
Projected Valuation: $400-500B at IPO

Roelof Botha noted their operational conservatism: "Remarkably sober minded and sensible and kind of mainstream in terms of the financial plan. You know, all along, they've stated they're gonna be more efficient, burn less, they've talked about converging early. It feels like they will very sensibly go public when they can, because in this kind of capital intensive business, I think at some point, there will be a strategic advantage to going public." [00:13:54]

Keith Rabois added growth context: "If anthropic's predicting the revenue is gonna triple next year...it's difficult to imagine, even in an amazing market like today, you're gonna get a much better multiple than in early 2026, right?" [00:15:40]

The panel's IPO price predictions: Jason 500B, Keith 500B, Roelof 350B, Harry 420B [00:16:11]

SpaceX

Description: Aerospace and satellite communications company
Why Mentioned: Secondary sale at $800B valuation tests limits of private market pricing
Metrics: $15B revenue, 30% growth, pursuing $800B valuation through secondary

Roelof Botha's analysis: "It's an amazing cut...It's actually amazing two companies. It's an amazing rocket company and then amazing communications and stalling company. It's doing $15 billion, growing plus or minus 30%. This year, that's a pretty hefty valuation. That's, you know, north of 40 times one rate for a company this year going 30." [00:02:02]

The Elon premium was acknowledged: "There's a huge amount of Elon, magic overlay and, you know, so far that magic has worked, but it's definitely a lot of non-obvious math baked into the price." [00:02:40]

Netflix

Description: Streaming entertainment platform
Why Mentioned: Case study of tech completely consuming traditional industry
Acquisition: Pursuing Warner Brothers Discovery for $82.7B

Roelof Botha's framing: "Netflix won, they ate the media industry, their market cap is $470 billion. The biggest studio is sub 200 billion dollars...The beauty of Netflix is, again, zooming out, because you have global distribution, you can monetize content better than anyone else. The value of the, you can pay more for content because you got more people to sell the content to. It's just that simple, right? The business model won." [00:19:30]

Airwallex

Description: Global payments and treasury platform, dual-headquartered in Australia and San Francisco
Why Mentioned: Extraordinary valuation discount despite comparable metrics to US competitors
Metrics: ~$1B ARR, raised $330M at $8B valuation led by Addition (significant Asia discount vs RAMP's $32B)

Keith Rabois framed the core question: "It's doing the same revenue that RAMP does...Let's assume it's growing the same as RAMP and has the same gross margins...It's trading at a quarter of the price...I literally don't know. It just seems a quarter of the price. It seems a bit of a wrap." [01:17:40]

The China data controversy emerged as the key risk, with Keith publicly raising concerns about data flows. Roelof Botha's investment stance: "If I was on the air wallet explored...I would put in an 8 billion and a part of my closing condition would be within 24 hours, we do not have a single paid employee in China. And we locate the entire service organization to some other jurisdiction." [01:27:22]

Sierra

Description: AI customer service platform
Why Mentioned: Example of enterprise AI implementation complexity creating moats

Harry Stebbings argued: "A lot of the defensibility, the mode, the skill is in the GTM and the implementation. And so I don't think it's just quite as simple as like, oh, well, like a new model will happen." [00:54:05]

Cursor & Replit

Description: AI-powered coding environments
Why Mentioned: Examples of AI apps using proprietary models and Chinese open-source alternatives

Keith Rabois: "So many of the high companies use their own LMS and their Chinese models, cursor's own model...So like all of our portfolio companies that are using their own LMS, they're not really using their own LMS, they're using cheap open source or if they could, they would use Chinese hosted models and sometimes they do because they're a fraction of the cost." [01:06:42]

4. People Identified

Chase Coleman (Tiger Global)

Description: Founder and Managing Partner of Tiger Global
Why Mentioned: For showing "love of the game" by raising $2.2B fund with 20% GP commit despite past challenges

Roelof Botha's assessment: "The guy who wanted that has got billions of dollars doesn't need the grace, right? It speaks to a love of the game that he says, I'm going to swallow my pride...and I was wrong. And I give him huge credit for not going, and now I'm just going to take my $2 billion, $5 billion, and stay in my house in Palm Beach or whatever." [00:12:57]

Harry Stebbings added: "I also think it pains him greatly that their brand has been tarnished in the LP community. And he will do everything possible to make our people money again. I think he is wounded and an incredible dude, actually. I'm a big fan of Chase." [00:36:01]

Naveen Rao (Databricks founder)

Description: Serial entrepreneur who founded Databricks
Why Mentioned: For raising $500M seed round at $5B valuation for new AI company based purely on founder track record

Roelof Botha referenced the classic phrase: "I can't remember who wrote the book...but it was the title was, once you're lucky twice, you're good. I'm Navine's done it twice. I mean, you just look at that and you go, at this point you can draw a line. And once you can draw a line, the VCs are gonna just show up on mass. It totally makes sense." [00:40:48]

Jack Zhang (Airwallex founder)

Description: Chinese-Australian founder of Airwallex
Why Mentioned: For building payments company at quarter the valuation of US competitors while navigating China controversy

Keith Rabois: "I don't know the CEO. He seems ultra-impressive from afar, not just from the recent press, but from following for a while in social media. You just, when this happens, you just have to be clear. You have to not, you have to not only say no US data sent to China, you have to make it unimpeachable." [01:25:06]

David Zaslav (Warner Brothers Discovery CEO)

Description: CEO of Warner Brothers Discovery
Why Mentioned: For clearly preferring Netflix acquisition over Paramount's hostile bid

Keith Rabois observed: "Clearly he does not want to do the paramount deal...It is interesting. And if you've been, for folks, if you've gone through M&A, you really want to have two real offers...But when you really have two, you really think about what you want to do with your life more. Just the interesting dynamic, because I just really don't think he wants to go work for Larry Jr. at Paramount." [01:24:18]

Marc Benioff (Salesforce CEO)

Description: Founder and CEO of Salesforce
Why Mentioned: For "LLMs are commodities" comment that sparked debate

Referenced in context of his hard disk drive analogy: "Mark actually used the analogy of it's like hard disk drives. And if you eat Clayton questions and it's the prototypical example of a widely competitive commoditized market where there were 30 providers." [01:03:12]

Keith Rabois defended Benioff's position through example: "Of course, Mark's right. Mark, just like Repplet today, can auto-rotate models without you knowing it. Salesforce is and should be doing the same thing. They have their own LLMs as they talked about." [01:09:16]

5. Operating Insights

The Three-Year Forward Pricing Limit Represents Peak Venture Optimism

Keith Rabois identified the historical boundary: "I think in 2021 we kind of reached peak of people paying three years ahead, right? In terms of multiples. Maybe we're back to that...Maybe three years ahead is the limit of VC can do even with the most optimist. You got to have, it has to tie to something when you put together the investment memo." [00:50:00]

This creates a disciplined framework for late-stage investing: investors can justify paying for future growth, but only within a bounded timeframe where projections maintain credibility. Beyond three years, the uncertainty overwhelms any analytical framework.

The "Loss of Memory" Switching Cost Is Real But Insufficient

In evaluating LLM defensibility, the panel debated memory as a moat. Harry argued: "I don't see the volatility of revenue on the consumer side because you have memory. And I think now, I don't know about you guys, but I'm seeing this so much in the other forums that I'm using where memory drives so much of the actual end product that I get. And that's incredibly sticky and we'll drive a lot of value." [01:06:16]

However, Keith challenged: "I think why Sam called the code red is imagine today, imagine today when we're doing this, if chatGPT went away tomorrow...Google AI mode's pretty good. Now it's not going to be our therapist, it's not going to have as much memory, but AI mode's pretty good...in a week, we'd be like whatever, we've moved on." [01:08:01]

The operational insight: memory creates friction but not lock-in at the consumer level. Enterprise applications have stronger moats through workflow integration.

Remove Strategic Objections Before They Become Sales Problems

Roelof Botha's framework for the Airwallex China exposure: "You will limit your pool of buyers dramatically if you have this significant exposure to this thing. How you get rid of that, you can talk about as a CEO. But it is one of the few things where you as a board member might have a little more perspective than someone. Because when you're down in it, you can go, that's silly. It's wrong...And one of the roles that a board member was to say, step back, dude, everything you're saying is true, but when it goes to ship no one will care." [01:30:00]

This sparked heated debate with Keith arguing for founder autonomy, but the core insight stands: perceived strategic risks must be eliminated proactively, regardless of whether they're technically valid concerns.

The Asset-Light Distribution Model Beats Asset-Heavy Production Every Time

From the Netflix-Warner Brothers discussion, Roelof Botha crystallized: "Because you have global distribution, you can monetize content better than anyone else. The value of the, you can pay more for content because you got more people to sell the content to. It's just that simple, right? The business model won." [00:27:45]

This extends beyond media to any industry where production can be separated from distribution. The company controlling distribution captures disproportionate value because they can monetize any production across their entire base, while producers are limited to their own distribution reach.

Dynamic Model Routing Is Table Stakes, Not Competitive Advantage

Keith Rabois described Replit's evolution: "Literally today, Repplet change...Now with Repplet, you don't pick your model anymore. It's all dynamic. It's all gone. You don't pick a model, you don't pick high reasoning, you don't pick short reasoning, you don't pick anything. It just, and it rotates Google and others." [01:06:23]

This represents a maturation of AI applications where model selection becomes invisible infrastructure rather than a user-facing decision. Companies that haven't implemented dynamic routing are operating at a disadvantage, as they can't optimize for cost, speed, and quality in real-time.

6. Overlooked Insights

The Stairstepped Valuation Round Structure Masks True Pricing Discovery

A critically important but briefly mentioned phenomenon: companies are raising multiple tranches at different prices simultaneously but only announcing the highest. Harry noted: "Although it's $505 billion on the sticker price when it's announced, there's actually several rounds before that Sequoia and Andrewson have done that are considerably cheaper and the $5 billion sticker seed is actually very different to the first two rounds before that no one's picked up on." [00:37:49]

Roelof Botha explained the mechanics: "You see one investor commit, and they commit some money to you at 300 million, some money at 500 million, and some money at a billion. And from the investor perspective, they make the commitment and even the checks at the same time. So their blended cost is five or six hundred million. So that's how they, that's their economics and that's the fact. But the entrepreneur can then announce a headline billion dollar raise." [00:37:24]

This insight is hugely significant because it means the "market clearing price" for hot deals is systematically misreported. The real price discovery happens in private tranches, while public announcements create FOMO at artificially high valuations. This explains why sophisticated investors can seemingly justify insane prices - their actual basis is 30-50% lower than reported.

The "85% Rule" of SaaS Growth Deceleration No Longer Applies to AI

Roelof Botha shared: "I think we published on this like 10, 15 years ago. In SAS, you could use this following rule of thumb with a high degree of accuracy. The growth rate for any year was roughly 85% of the prior years growth." [00:48:12]

This rule-of-thumb allowed precise long-term revenue modeling in SaaS. Its breakdown in AI companies represents a fundamental shift that hasn't been widely discussed. If AI companies decelerate faster (10X to 3X to 1.5X), many current valuations are severely wrong. If they decelerate slower (10X to 6X to 3X), we're still dramatically underpricing winners. The uncertainty in this deceleration rate is perhaps the single most important unknown in venture capital today, yet it's rarely explicitly discussed in these terms.

The implication: investors operating with SaaS mental models will systematically misprice AI companies, but we don't yet have enough longitudinal data to build the new model. The winners will be those who correctly predict the new deceleration curve.